Basic Text Analysis in R

Basic Text Analysis in R: WAIAI on-demand workshop

Introduction

What is text analysis?

Text analysis is quantitative analysis that is used to make unstructured text, such as reviews, open-ended survey responses, or customer tickets, usable for analysis in a range of settings. Text analysis results are typically observational, but can provide useful exploratory insights and support hypothesis generation.

How can basic text analysis be used to provide insight?

1. Define question to be addressed

2. Obtain text data after evaluating data sources, quality, and any ethical considerations

3. Conduct exploratory data analyses

4. Apply text analysis techniques, such as (but not limited to):

• Sentiment analysis for quantifying positive and negative linguistic sentiment

• Term frequency and keyword extraction to examine important terms

• Topic modeling or cluster analyses to summarize themes

• Text classification to categorize text into labels of interest

5. Interpret and visualize results

6. Formulate insights

Load packages

See all software references at the end of the tutorial.

#load packages and comment their uses
library(tidyverse) #data cleaning, organization, and visualization
library(psych) #summarize descriptive statistics and distributions
library(stringr) #data cleaning, organization, and visualization
library(skimr) #summarize descriptive statistics and distributions
library(tidytext) #sentiment analysis, text cleaning, and word frequency
library(textclean) #text cleaning and word frequency
library(wordcloud) #text cleaning and word frequency
library(tm) #word association
library(vader) #sentiment analysis
library(topicmodels) #topic modeling
library(MetBrewer) #plot color palettes

# color palette
MetBrew_Egypt <- MetBrewer::met.brewer("Egypt", n = 5)
MetBrew_Tam <- MetBrewer::met.brewer("Tam", n = 15)

Text data cleaning and exploration

Data source: random subset of 10,000 reviews (2% of dataset) from the Amazon Fine Foods reviews dataset (McAuley & Leskovec, 2013), learn more and download data here: https://snap.stanford.edu/data/web-FineFoods.html

#read data downloaded from Amazon Fine Foods reviews source
df_raw <- read.delim("../foods.txt", header = F) #load raw data saved to wd
df_raw$V1 <- iconv(df_raw$V1, from = "", to = "UTF-8") #convert all text to UTF-8

#cleaning raw dataframe to create one column per item within V1, one row per review using regex and dplyr
df <- df_raw %>%
  dplyr::mutate(V1 = str_split(V1, "\\n")) %>% 
  unnest(V1) %>%
  dplyr::mutate(V1 = str_trim(V1)) %>%
  dplyr::filter(V1 != "") %>%
  dplyr::mutate(V1 = str_split(V1, "(?=review/) | (?=product/)")) %>%
  unnest(V1) %>%
  dplyr::mutate(V1 = str_trim(V1)) %>%
  dplyr::filter(V1 != "")  %>%
  dplyr::mutate(names = str_extract(V1, "^review/\\w+|^product/\\w+"),
         values = str_remove(V1, "^review/\\w+:\\s*|^product/\\w+:\\s*")) %>%
  select(names, values) %>%
  dplyr::mutate(review_id = as.numeric(cumsum(names == "product/productId"))) %>%
  pivot_wider(names_from = names, values_from = values) %>%
  select(-c("review/time", "review/summary", "review/profileName")) %>%
  rename("productId" = "product/productId",
         "userId" = "review/userId",
         "score" = "review/score",
         "text" = "review/text") %>%
    dplyr::mutate(userId = match(userId, sample(unique(userId))),
                productId = match(productId, sample(unique(productId)))) %>%
  select(c("review_id", "productId", "userId", "score", "text")) %>%
  as.data.frame() 

#save new ids as categorical variables
df$review_id <- as.factor(df$review_id)
df$userId <- as.factor(df$userId)
df$productId <- as.factor(df$productId)

rm(df_raw) #remove giant raw dataset from environment

set.seed(22) #set seed to sample reproducibly
df <- df %>% 
  dplyr::slice_sample(n = 10000) %>% #take random set of 10000 reviews - 8 rows each
  unnest(c(2:5)) #unnest list cols

#replace "NULL" with NA
df[df=="NULL"] <- NA

#write.csv(df, "./foods_small.csv") ## optional: save small top 10000 reviews for future use
#df <- read.csv("./foods_small.csv") %>% select(-c(X)) #optional: read in saved data to save future pre-processing time

#examine text data
knitr::kable(skim(df)) #overview of data

skim_type	skim_variable	n_missing	complete_rate	character.min	character.max	character.empty	character.n_unique	character.whitespace	factor.ordered	factor.n_unique	factor.top_counts
character	score	3	0.9997000	3	3	0	5	0	NA	NA	NA
character	text	19	0.9981002	33	6839	0	9814	0	NA	NA	NA
factor	review_id	0	1.0000000	NA	NA	NA	NA	NA	FALSE	10000	311: 2, 4: 1, 16: 1, 45: 1
factor	productId	0	1.0000000	NA	NA	NA	NA	NA	FALSE	3525	358: 82, 949: 80, 756: 76, 682: 72
factor	userId	0	1.0000000	NA	NA	NA	NA	NA	FALSE	9280	484: 7, 495: 7, 527: 7, 165: 6

Takeaway: We should have 5 columns in the “df” dataframe that represent a random subset of 10,000 Amazon Fine Foods review data that we will be using for this tutorial:

• review_id (factor): review identifier (created during data cleaning based on row number)

• productId (factor): product identifier for product being reviewed (replaced with sequential number ID during data cleaning)

• userId (factor): reviewer identifier (replaced with sequential number ID during data cleaning)

• score (character): number of stars given by the reviewer

• text (character): complete review text

Question to be addressed: What are some common customer experiences and painpoints with Amazon Fine Foods products?

Exploratory data analysis

Distribution of review scores

#descriptive stats
psych::describe(df$score) #describe scores

##     vars    n mean   sd median trimmed mad min max range  skew kurtosis   se
## X1*    1 9998 4.14 1.32      5    4.41   0   1   5     4 -1.36     0.46 0.01

#plot scores
df %>% 
  drop_na(score) %>% 
  ggplot(aes(x = as.factor(score))) +
  geom_bar(color = "darkgray", fill = "gray") +
  labs(title = "Frequency of review scores", 
       x = "Score", y = "Frequency") +
  theme_classic()

Takeaway: Most reviews have 5 stars.

Average number of reviews per product

df %>% 
  count(productId) %>% 
  psych::describe() #describe number of reviews per product

##            vars    n    mean      sd median trimmed     mad min  max range skew
## productId*    1 3525 4867.68 2815.71   4853 4861.48 3660.54   1 9768  9767 0.01
## n             2 3525    2.84    5.76      1    1.62    0.00   1   82    81 6.93
##            kurtosis    se
## productId*    -1.20 47.43
## n             62.69  0.10

#plot number of reviews per product
df %>% 
  count(productId)  %>%
  ggplot(aes(x = n)) +
  geom_histogram(color = "darkgray", fill = "gray", binwidth = 5) +
  labs(title = "Frequency of reviews per product", 
       x = "Review count per product", y = "Frequency") +
  theme_classic()

Takeaway: Most products have several reviews.

Comparing review scores of highly-reviewed and non-highly-reviewed products

#examine whether review scores vary among highly reviewed and non-highly-reviewed products
df <- df %>% 
  add_count(productId, name = "n_reviews") %>% #add column with count for number of reviews 
  dplyr::group_by(productId) %>%
  dplyr::mutate(highly_reviewed = ifelse(n_reviews > 2.84, "highly-reviewed", "not highly-reviewed")) %>% #create new column denoting if product is above/below mean of n reviews
  dplyr::ungroup() 

psych::describeBy(df$score, df$highly_reviewed) #examine whether scores differ on number of reviews

## 
##  Descriptive statistics by group 
## group: highly-reviewed
##     vars    n mean   sd median trimmed mad min max range  skew kurtosis   se
## X1*    1 6684 5.13 1.29      6    5.39   0   1   6     5 -1.33     0.47 0.02
## ------------------------------------------------------------ 
## group: not highly-reviewed
##     vars    n mean   sd median trimmed mad min max range skew kurtosis   se
## X1*    1 3314 4.15 1.38      5    4.44   0   1   5     4 -1.4     0.44 0.02

#plot scores by high/low review
df %>% 
  drop_na(score) %>%
  ggplot(aes(x = as.factor(score))) +
  geom_bar(color = "darkgray", fill = "gray") +
  labs(title = "Frequency of review scores for highly-reviewed and not-highly-reviewed products",
       x = "Score", y = "Frequency") +
  theme_classic() +
  facet_wrap(~highly_reviewed)

Takeaway: Products that received more reviews than the mean and products that received fewer don’t differ substantially in terms of the distributions of stars given.

Average number of reviews per reviewer

df %>% 
  count(userId) %>% 
  psych::describe() #describe number of reviews per reviewer

##         vars    n     mean       sd median  trimmed      mad min   max range
## userId*    1 9280 28765.37 16478.39  28659 28773.47 21057.37   9 57220 57211
## n          2 9280     1.08     0.35      1     1.00     0.00   1     7     6
##         skew kurtosis     se
## userId* 0.01    -1.20 171.06
## n       6.50    60.94   0.00

#plot number of reviews per reviewer
df %>% 
  count(userId)  %>%
  ggplot(aes(x = n)) +
  geom_histogram(color = "darkgray", fill = "gray", binwidth = 1) +
  labs(title = "Frequency of reviews per reviewer", 
       x = "Review count per reviewer", y = "Frequency") +
  theme_classic()

Takeaway: Most reviewers reviewed only once in this subset.

Examine wordcount

#add wordcount column
df <- df %>%
  as.data.frame() %>%
  dplyr::group_by(review_id) %>%
  dplyr::mutate(review_wordcount = str_count(text, pattern = "\\w+")) %>% #add wordcount column using regex
  dplyr::ungroup() 
df$review_wordcount <- as.numeric(df$review_wordcount)

#examine results
psych::describe(df$review_wordcount) #describe wordcount

##    vars    n  mean   sd median trimmed mad min  max range skew kurtosis   se
## X1    1 9982 84.64 82.7     60   68.95  43   6 1318  1312 3.87     27.9 0.83

hist(df$review_wordcount, breaks = 100) #wordcount histogram

Takeaway: Review length is relatively short.

#plot review score and review length
df %>%
  subset(review_wordcount <= (84.64+(3*82.70))) %>% #subset to remove review length outliers more than 3 SD from average review length
  ggplot(aes(y = as.numeric(review_wordcount), x = as.factor(score), color = as.factor(score), fill = as.factor(score))) +
  geom_jitter(size = 1, color = "gray") +
  geom_violin(alpha = 0.7) +
  scale_fill_manual(values = MetBrew_Egypt) +
  scale_color_manual(values = MetBrew_Egypt) +
  geom_boxplot(width = 0.1, color = "white") +
  labs(x = "Review Score", y = "Review wordcount") +
  theme_classic() +
  theme(legend.position = "none") 

Takeaway: Review scores do not appear to vary much with length.

Text cleaning

• Some analyses suggest text cleaning to reduce noise and improve analysis accuracy and efficiency, and should be used only when recommended.

• Here, we will clean text by placing it in lowercase, removing extra whitespace, and replacing numeric characters with words (e.g., “1” -> “one”).

• We will also create a version of the text data with further cleaning that is needed for some analyses, with:

  • punctuation removed and contractions replaced
  • lemmatization (i.e., changed to its morphological base, so “running” and “ran” become “run”).
  • stopwords removed (i.e., common words that do not often have universal meaning in some text analyses, e.g., “the,” “a”), and
  • tokenization (i.e., one token per row; here, one word per row)

• Refer to best practices and documentation for specific analyses on which text cleaning steps are recommended (if any).

  • Consider whether an analysis is compositional (i.e., takes context into account vs. considering each word individually) before implementing text cleaning steps.

  • Sensitivity analyses can be a great tool to examine whether results change with and without text cleaning.

• Be transparent about text cleaning steps taken when sharing results.

#placing all text in lowercase, replacing numbers with numeric in text, and removing extra white space to clean text up
df$text <- tolower(df$text) #place all text in lowercase
df$text <- trimws(df$text) #remove extra whitespace
df$text <- textclean::replace_number(df$text, remove = T) #replace numeric characters
df$text <- gsub("[0-9]+", "", df$text) #remove any remaining numbers altogether
df$text <- gsub("<[^>]*>", "", df$text) #remove css tags contained in < >

#see results
knitr::kable(head(df[, c("review_id", "text", "review_wordcount")], n = 5))

review_id	text	review_wordcount
24421	i have never been much of a soup person, unless down with a cold or too tired to cook up something fancy. once heated, i crumbled tortilla chips on top (or even some shredded cheese). since i love mexican food, this satisfied me for the evening. i was advised by progresso that they have discontinued this soup. thank goodness amazon had some for purchasing.	64
39808	the bars make for a nice snack and they taste natural but are not something that i would eat for pleasure.	21
56240	we really enjoyed this product – containes no sugar (contains dates and such) and it’s healthy while giving you the chocolate fix you need!	24
70928	the toy is excellent. my golden retriever is teething and this has been perfect for her.-stars because product description is not accurate. picture showed t-rex, but i received a brontosaurus. no big deal but if there are multiple models the description should indicate which one you’re getting	53
77000	type of pasta: orzo, it resembles more like rice than pasta, so this dish felt like a chicken,cheesy rice,broccoli type of dish as opposed to straight pasta.resembles: hamburger helper’s but much cheesierdifficulty: very easytime: minutesa box fills up: about people (maybe more if you eat less)amount of broccoli: very little, wish they had more! but yes, very little, though you can taste it. basically very small little flecks of freeze dried broccoli.amount of cheese: good amount of cheese for people!taste: b+ease to make: anutrition: c+/b-overall: b	120

df_words <- df #save separate lemmatized df
df_words$text <- textclean::replace_contraction(df_words$text) #replace contractions
df_words$text <- gsub("[[:punct:]]", "", df_words$text) #remove remaining punctuation
df_words$text <- textstem::lemmatize_strings(df_words$text) #lemmatize words

#tokenize
df_words <- df_words %>% 
  tidytext::unnest_tokens(word, text, token = "words") %>% 
  dplyr::filter(!is.na(word))

stopwords <- subset(stop_words, lexicon == "snowball") #select stopword lexicon

#remove stopwords from data
df_words <- df_words %>% 
  dplyr::filter(!word %in% stopwords$word) %>% 
  dplyr::filter(!is.na(word)) 

#see results
knitr::kable(head(df_words[, c("review_id", "word")], n = 5)) #top 5 rows

review_id	word
24421	never
24421	much
24421	soup
24421	person
24421	unless

Basic text analysis

Examining frequent words and keywords

Most frequent words and bigrams

Most frequent words

#frequent word count dataframe
df_count <- df_words %>%
  count(word, sort = TRUE)

#show top frequent words table
knitr::kable(head(df_count, n = 25))

word	n
good	7249
like	5261
much	4943
can	4746
taste	4636
flavor	3735
get	3336
one	3335
will	3214
love	3181
product	3121
make	3113
just	3086
try	2999
use	2901
great	2838
buy	2754
coffee	2742
food	2660
tea	2454
find	2339
dog	2317
eat	2297
little	2007
go	1883

## frequent words wordcloud
wordcloud(df$text, min.freq = 500, colors = brewer.pal(12, "Dark2"))

#compare most frequent words among high and low scored reviews
df_stars <- df %>%
  dplyr::select(c(review_id, score)) %>%
  dplyr::mutate(high_low_score = ifelse(score == "1.0" | score == "2.0", "low",
                                 ifelse(score == "5.0", "high", NA))) #create high/low review score column

#plot top 25 words in high and low score reviews
df_words %>%
  left_join(., df_stars, by = c("review_id")) %>%
  subset(!is.na(high_low_score)) %>%
  dplyr::group_by(high_low_score) %>%
  count(word, sort = TRUE) %>% 
  top_n(n = 25) %>% 
  dplyr::mutate(word = reorder(word, n)) %>% 
  dplyr::ungroup() %>%
    ggplot(aes(n, reorder_within(word, n, high_low_score))) + 
    geom_col(color = "gray", fill = "darkgray") +
    labs(y = "Word", x = "Frequency ", title = "Amazon fine foods reviews subset top 25 most frequent words",
         subtitle = "Among high-scoring (5 star) and low-scoring (1 or 2 star) reviews") +
    geom_text(aes(label = n), hjust = 1, colour = "white") +
    theme_classic() +
    facet_wrap(~high_low_score, scales = "free")

## Selecting by n

Takeaway: The top few words across all reviews are “good,” “like,” “much,” “can,” and “taste”. 5-star reviews and 1- and 2-star reviews feature fairly similar words, although each word is taken out of context here.

Sensitivity analysis: Compare to most frequent words without text cleaning

#frequent word count with raw text
df_count2 <- df %>%
  tidytext::unnest_tokens(word, text, token = "words") %>% #redo tokenization on raw text without cleaning
  dplyr::filter(!is.na(word)) %>%
  count(word, sort = TRUE) #count occurrences of raw words

#top frequent words table
knitr::kable(head(df_count2, n = 25)) #show top 25 words

word	n
the	33258
i	25459
and	22868
a	21338
to	18097
it	15898
of	14419
is	13309
this	11731
for	9676
in	9596
my	7815
that	7647
but	6761
with	6308
not	6094
have	6089
you	5882
are	5803
was	5677
they	5309
as	4996
like	4757
on	4725
so	4536

Most frequent bigrams

#frequent bigram count
df_bigrams <- df %>%
  unnest_tokens(bigram, text, token = "ngrams", n = 2) %>% #repeat tokenization where token = bigram
  dplyr::filter(!is.na(bigram)) %>%
  tidyr::separate(bigram, c("word1", "word2"), sep = " ") %>%
  dplyr::filter(!word1 %in% stopwords$word,
         !word2 %in% stopwords$word)  #remove stopwords

df_bigrams$word1 <- textstem::lemmatize_strings(df_bigrams$word1) #lemmatize words 
df_bigrams$word2 <- textstem::lemmatize_strings(df_bigrams$word2) #lemmatize words

#paste both words in bigram together in new column
df_bigrams <- df_bigrams %>%
  dplyr::mutate(bigram = paste(word1, word2, sep = " ")) %>%
  select(-c(word1, word2)) %>%
  count(bigram, sort = TRUE) 

#top frequent bigrams table
knitr::kable(head(df_bigrams, n = 25))

bigram	n
taste like	459
k cup	425
highly recommend	322
dog food	321
green tea	319
gluten free	317
grocery store	294
peanut butter	238
much good	233
dog love	232
year old	230
taste good	217
taste great	211
really like	203
dark chocolate	202
cat food	176
really good	175
can get	171
great taste	171
great product	158
just like	154
potato chip	154
good price	152
look like	141
good taste	139

#compare high and low scored reviews
bigram_df_plot <- df %>%
  left_join(., df_stars, by = c("review_id")) %>%
  subset(!is.na(high_low_score)) %>%
  dplyr::group_by(high_low_score) %>%
  unnest_tokens(bigram, text, token = "ngrams", n = 2) %>%
  dplyr::filter(!is.na(bigram)) %>%
  tidyr::separate(bigram, c("word1", "word2"), sep = " ") %>%
  dplyr::filter(!word1 %in% stopwords$word,
         !word2 %in% stopwords$word)  #remove stopwords

bigram_df_plot$word1 <- textstem::lemmatize_strings(bigram_df_plot$word1) #lemmatize words
bigram_df_plot$word2 <- textstem::lemmatize_strings(bigram_df_plot$word2) #lemmatize words

#plot top bigrams in high and low scoring reviews
bigram_df_plot %>%
  dplyr::mutate(bigram = paste(word1, word2, sep = " ")) %>%
  select(-c(word1, word2)) %>%
  count(bigram, sort = TRUE) %>% 
  top_n(n = 25) %>% 
  dplyr::mutate(bigram = reorder(bigram, n)) %>% 
  dplyr::ungroup() %>%
    ggplot(aes(n, reorder_within(bigram, n, high_low_score))) + 
    geom_col(color = "gray", fill = "darkgray") +
    labs(y = "Bigram", x = "Frequency ", title = "Amazon fine foods reviews top 25 most frequent bigrams",
         subtitle = "Among high-scoring (5 star) and low-scoring (1 or 2 star) reviews") +
    geom_text(aes(label = n), hjust = 1, colour = "white") +
    theme_classic() +
    facet_wrap(~high_low_score, scales = "free")

## Selecting by n

Takeaway: The top bigrams words across all reviews are “gluten free,” “green tea,” “dog food,” “highly recommend,” and “k cup”. 5-star and 1-and 2-star reviews feature some of these (e.g., “k cup,” “gluten free,” “dog food,” “green tea”), while 5-star reviews feature some more positive bigrams (“highly recommend,” “great product,” “really good,” “taste great”) 1- and 2-star reviews feature some more negative bigrams (“taste like,” “look like,” “never buy”).

Sensitivity analysis: Compare to most frequent bigrams without text cleaning

#frequent bigrams with raw text
df_bigrams2 <- df %>%
  unnest_tokens(bigram, text, token = "ngrams", n = 2) %>% #tokenize bigrams without further cleaning
  dplyr::filter(!is.na(bigram)) %>%
  tidyr::separate(bigram, c("word1", "word2"), sep = " ") %>%
  dplyr::mutate(bigram = paste(word1, word2, sep = " ")) %>%
  select(-c(word1, word2)) %>%
  count(bigram, sort = TRUE)

#top frequent bigrams table
knitr::kable(head(df_bigrams2, n = 25))

bigram	n
of the	2788
in the	2330
i have	2143
it is	2055
this is	2029
is a	1662
i was	1482
they are	1459
and i	1370
if you	1363
on the	1329
and the	1174
this product	1170
for a	1140
to be	1140
it was	1130
i am	1090
to the	1068
for the	1048
in a	1033
is the	974
the best	970
but i	926
a little	908
i would	895

Examine most frequent words via Document-Term Matrix and correlated terms

#create document-term matrix from tf-idf
df_dtm <- df_words %>%
  count(review_id, word, sort = TRUE) %>% #count words per review
  tidytext::bind_tf_idf(term = 'word',
              document = 'review_id',
              n = 'n') %>% #add term frequency info
  dplyr::filter(!is.na(tf)) %>%
  dplyr::select(review_id, word, n) %>%
  pivot_wider(id_cols = "review_id", names_from = "word", values_from = "n") %>% #pivot wider
  column_to_rownames(var='review_id') %>%
  as.data.frame()

#replace NA with 0 for DTM
df_dtm[is.na(df_dtm)] <- 0

#show top few rows to show example of DTM
knitr::kable(head(df_dtm[,c(1:20)], n = 5)) #select top few words in DTM 

	thai	tea	mg	dog	coffee	food	can	pad	almond	noodle	rise	cat	gopher	chip	oil	green	capsule	fructose	treat	juice
53927	53	0	0	0	0	3	3	24	0	10	0	0	0	0	1	3	0	0	0	0
36094	0	37	2	0	1	0	0	0	0	0	0	0	0	0	0	18	0	0	0	0
59678	0	0	32	0	0	2	2	0	0	0	0	0	0	0	0	0	0	1	0	2
71779	0	0	0	32	0	28	4	0	0	0	0	1	0	0	0	0	0	0	1	0
36306	0	1	0	0	29	1	3	0	0	0	0	0	0	0	0	1	0	0	0	0

#cast DTM using tidytext cast_dtm
data_dtm <- df_words %>% 
  dplyr::count(review_id, word) %>% #count words per review
  dplyr::mutate(review_id = as.numeric(review_id)) %>%
  tidytext::cast_dtm(document = review_id, term = word, value = n) #cast document term matrix

#find frequent terms with over 1000 uses using tm package
findFreqTerms(data_dtm, 1000)

##  [1] "find"      "flavor"    "get"       "good"      "look"      "make"     
##  [7] "order"     "love"      "package"   "really"    "eat"       "just"     
## [13] "like"      "much"      "store"     "taste"     "add"       "little"   
## [19] "price"     "also"      "can"       "chocolate" "first"     "go"       
## [25] "know"      "one"       "sugar"     "think"     "time"      "treat"    
## [31] "use"       "will"      "year"      "great"     "dog"       "food"     
## [37] "give"      "try"       "buy"       "mix"       "recommend" "day"      
## [43] "tea"       "come"      "now"       "product"   "bite"      "coffee"   
## [49] "want"      "box"       "say"       "amazon"    "drink"     "even"     
## [55] "cup"       "bag"

#which words are correlated with "find"?
findAssocs(data_dtm, c("find"), (0.15))

## $find
##    can  store   much   hard amazon 
##   0.20   0.18   0.16   0.16   0.16

Takeaway: Across all reviews, the most common terms by word frequency are “find,” “flavor,” “get,” “good,” and “look”. “Find” is most correlated with “will,” “store,” “much,” “cheap,” and “amazon”.

#frequent and associated terms for 5-star reviews
data_dtm_high <- df_words %>% 
  subset(score == "5.0") %>%
  dplyr::count(review_id, word) %>%
  dplyr::mutate(review_id = as.numeric(review_id)) %>%
  tidytext::cast_dtm(document = review_id, term = word, value = n)

findFreqTerms(data_dtm_high, 500)

##  [1] "love"      "package"   "really"    "brand"     "eat"       "find"     
##  [7] "good"      "just"      "keep"      "like"      "much"      "store"    
## [13] "taste"     "buy"       "food"      "mix"       "recommend" "will"     
## [19] "day"       "give"      "tea"       "try"       "add"       "cat"      
## [25] "enjoy"     "get"       "water"     "year"      "also"      "can"      
## [31] "one"       "flavor"    "little"    "favorite"  "come"      "now"      
## [37] "order"     "product"   "work"      "great"     "make"      "go"       
## [43] "look"      "take"      "time"      "sweet"     "use"       "coffee"   
## [49] "want"      "first"     "think"     "amazon"    "drink"     "even"     
## [55] "pack"      "price"     "purchase"  "healthy"   "sugar"     "box"      
## [61] "know"      "say"       "cup"       "delicious" "chocolate" "chip"     
## [67] "bag"       "snack"     "need"      "dog"       "treat"

#which words are correlated with "find"?
findAssocs(data_dtm_high, c("find"), (0.15))

## $find
##    hard   store     can  amazon  appall finally grocery    much    year   local 
##    0.22    0.21    0.21    0.21    0.18    0.17    0.16    0.16    0.16    0.16 
## process  search    food 
##    0.16    0.16    0.15

Takeaway: The most common terms in 5-star reviews by word frequency are “love,” “package,” “really,” “brand,” and “eat”. “Find” is most correlated with “hard,” “can,” “store,” “amazon,” and “appall”.

#frequent and associated terms for 1 or 2-star reviews
data_dtm_low <- df_words %>% 
  subset(score == "1.0" | score == "2.0") %>%
  dplyr::count(review_id, word) %>%
  dplyr::mutate(review_id = as.numeric(review_id)) %>%
  tidytext::cast_dtm(document = review_id, term = word, value = n)

findFreqTerms(data_dtm_low, 300)

##  [1] "flavor"  "get"     "good"    "make"    "order"   "can"     "eat"    
##  [8] "go"      "like"    "much"    "one"     "taste"   "think"   "use"    
## [15] "will"    "just"    "buy"     "dog"     "food"    "product" "try"    
## [22] "even"    "coffee"

#which words are correlated with "find"?
findAssocs(data_dtm_low, c("find"), (0.15))

## $find
##            casei        dependant       physiology            react 
##             0.24             0.24             0.24             0.24 
##         rockstar themselvesupdate             tryi            annes 
##             0.24             0.24             0.24             0.24 
##            apear        guideline         honolulu          hygiene 
##             0.24             0.24             0.24             0.24 
##           inwhat        marketsit         regimine       substiture 
##             0.24             0.24             0.24             0.24 
##             usai             yoga          midwest           person 
##             0.24             0.24             0.23             0.21 
##           entire       reasonably           pickup            thier 
##             0.21             0.21             0.21             0.20 
##          success           seldom              try   challengecliff 
##             0.20             0.20             0.19             0.19 
##          deviate           dietas              ibs             road 
##             0.19             0.19             0.19             0.19 
##            rural       walmartcom       weightthis         strength 
##             0.19             0.19             0.19             0.19 
##      unavailable               cv     increasingly        tastewell 
##             0.19             0.19             0.19             0.19 
##           effect             much          request           review 
##             0.18             0.18             0.18             0.17 
##      restriction       substitute             bite          product 
##             0.17             0.17             0.16             0.16 
##             safe             line              lie             easy 
##             0.16             0.16             0.16             0.16 
##            towel             wary             youi          patient 
##             0.16             0.16             0.16             0.16 
##      outstanding            yield             size           severe 
##             0.16             0.16             0.15             0.15

Takeaway: The most common terms in 1 and 2-star reviews by word frequency are “get,” “good,” “flavor,” “make,” and “order”. “find” is most correlated with “casei,” “dependant,” and “physiology”, “react”, and “rockstar”.

Topic model (LDA)

We can summarize the topics discussed across all reviews using Latent Dirichlet Allocation (LDA) topic modeling (Blei, Ng, & Jordan, 2003). There are many approaches for topic modeling but LDA is a common one and effective to provide an overview of review content here.

#estimate an LDA topic model with 15 topics
set.seed(2025)
reviews_lda <- topicmodels::LDA(data_dtm, k = 15, list = control(seed = 2025)) #run topic model with 15 topics

#get top 15 terms per topic
knitr::kable(topicmodels::get_terms(reviews_lda, 15))

Topic 1	Topic 2	Topic 3	Topic 4	Topic 5	Topic 6	Topic 7	Topic 8	Topic 9	Topic 10	Topic 11	Topic 12	Topic 13	Topic 14	Topic 15
product	like	good	flavor	food	organic	tea	dog	order	sugar	chocolate	price	almond	coffee	use
use	taste	make	chip	cat	ingredient	green	treat	box	drink	bar	amazon	salt	cup	hot
cereal	can	use	good	eat	product	good	love	product	taste	good	buy	nut	good	sauce
good	good	cook	like	dog	fat	flavor	get	package	flavor	taste	store	seed	flavor	popcorn
hair	much	mix	taste	can	list	taste	chew	arrive	juice	like	good	blue	like	good
like	try	like	eat	much	food	like	one	good	fruit	flavor	find	good	taste	make
bottle	just	noodle	snack	good	much	drink	will	receive	like	make	can	snack	much	oz
shampoo	get	just	love	like	contain	much	give	cookie	water	much	product	healthy	one	just
much	think	taste	great	love	good	bag	good	one	sweet	will	order	much	strong	taste
dry	really	can	much	one	make	ginger	tooth	will	can	dark	much	taste	try	will
oatmeal	one	much	try	year	fiber	try	like	make	much	just	great	diamond	roast	much
will	bad	pasta	bag	old	protein	make	much	time	good	eat	get	blood	drink	add
oil	say	easy	peanut	get	rice	love	small	can	add	mix	love	keep	blend	spice
work	will	free	butter	baby	g	cup	can	gift	soda	love	ship	great	make	get
try	give	gluten	potato	will	gram	one	great	love	use	great	purchase	eat	bean	pop

#visualize top 5 words per topic
tidytext::tidy(reviews_lda, matrix = "beta") %>%
  dplyr::group_by(topic) %>%
  top_n(5, beta) %>% #count words with highest beta per topic
  dplyr::ungroup() %>%
  ggplot(aes(x = beta, y = reorder_within(term, beta, topic), fill = as.factor(topic))) +
  geom_col() +
  scale_fill_manual(values = MetBrew_Tam) +
  scale_color_manual(values = MetBrew_Tam) +
  facet_wrap(~ topic, scales = "free") +
  labs(title = "Top 5 words per LDA topic across Amazon fine food reviews subset",
       x = "Probability of word belonging to topic (beta)", y = "Word") +
  theme_classic() +
  theme(legend.position = "none")

Takeaway: This subset of reviews is primarily about specific products or broader consumer experiences such as finding items, pricing, and shipping.

Sensitivity analysis: Compare to LDA done without text cleaning

#redoing DTM as shown above without text cleaning prior to tokenization
data_dtm2 <- df %>%
  tidytext::unnest_tokens(word, text, token = "words") %>% 
  dplyr::filter(!is.na(word)) %>%
  dplyr::count(review_id, word) %>%
  dplyr::mutate(review_id = as.numeric(review_id)) %>%
  tidytext::cast_dtm(document = review_id, term = word, value = n)

#estimate an LDA topic model with 15 topics
reviews_lda2 <- topicmodels::LDA(data_dtm2, k = 15, list = control(seed = 2025)) #run topic model

#get top 15 terms per topic
knitr::kable(topicmodels::get_terms(reviews_lda2, 15))

Topic 1	Topic 2	Topic 3	Topic 4	Topic 5	Topic 6	Topic 7	Topic 8	Topic 9	Topic 10	Topic 11	Topic 12	Topic 13	Topic 14	Topic 15
a	to	i	the	the	i	a	the	the	the	it	the	and	to	a
of	a	a	to	a	and	i	and	i	i	to	and	i	i	i
i	and	it	a	it	the	of	of	and	to	and	this	is	it	this
it	of	in	i	and	a	is	to	is	it	the	in	my	is	the
to	i	and	and	my	is	them	a	to	is	of	it	not	this	in
is	it	this	of	for	but	the	for	for	and	this	of	to	for	but
the	this	the	in	to	it	for	are	this	a	i	you	of	of	of
and	for	my	are	in	so	are	is	was	my	my	with	that	the	it
was	not	but	for	this	have	these	have	a	you	was	not	it	with	with
for	that	is	that	on	they	my	in	of	but	that	a	in	in	as
you	with	of	with	but	these	was	i	my	in	for	i	they	that	is
with	you	to	as	that	are	they	on	you	so	them	flavor	like	and	have
tea	food	that	so	with	that	and	with	that	that	is	is	you	these	so
but	in	like	have	like	to	so	you	like	at	but	as	this	as	for
in	dog	for	not	they	this	have	good	one	like	in	was	these	they	and

Basic sentiment analysis

Using a compositional method (Valence Aware Dictionary and sEntiment Reasoner, i.e., VADER; Hutto & Gilbert, 2014)

#get VADER compound sentiment scores for each review using minimally-cleaned text data since it is compositional
df_vader <- df %>%
  dplyr::group_by(review_id) %>%
  dplyr::mutate(vader_output = vader::vader_df(text)) %>% #use vader package to assign sentiment values per review
  dplyr::mutate(sentiment_vader = vader_output$compound) %>% #create new column for compound vader sentiment score
  select(-c(vader_output)) %>% #remove extraneous vader output
  dplyr::ungroup() %>%
  as.data.frame()

#examine sentiment
knitr::kable(skim(df_vader))

skim_type	skim_variable	n_missing	complete_rate	character.min	character.max	character.empty	character.n_unique	character.whitespace	factor.ordered	factor.n_unique	factor.top_counts	numeric.mean	numeric.sd	numeric.p0	numeric.p25	numeric.p50	numeric.p75	numeric.p100	numeric.hist
character	score	3	0.9997000	3	3	0	5	0	NA	NA	NA	NA	NA	NA	NA	NA	NA	NA	NA
character	text	19	0.9981002	33	6608	0	9814	0	NA	NA	NA	NA	NA	NA	NA	NA	NA	NA	NA
character	highly_reviewed	0	1.0000000	15	19	0	2	0	NA	NA	NA	NA	NA	NA	NA	NA	NA	NA	NA
factor	review_id	0	1.0000000	NA	NA	NA	NA	NA	FALSE	10000	311: 2, 4: 1, 16: 1, 45: 1	NA	NA	NA	NA	NA	NA	NA	NA
factor	productId	0	1.0000000	NA	NA	NA	NA	NA	FALSE	3525	358: 82, 949: 80, 756: 76, 682: 72	NA	NA	NA	NA	NA	NA	NA	NA
factor	userId	0	1.0000000	NA	NA	NA	NA	NA	FALSE	9280	484: 7, 495: 7, 527: 7, 165: 6	NA	NA	NA	NA	NA	NA	NA	NA
numeric	n_reviews	0	1.0000000	NA	NA	NA	NA	NA	NA	NA	NA	14.5380462	19.0767863	1.000	2.000	5.000	22.000	82	▇▂▁▁▁
numeric	review_wordcount	19	0.9981002	NA	NA	NA	NA	NA	NA	NA	NA	84.6415548	82.7039978	6.000	35.000	60.000	103.000	1318	▇▁▁▁▁
numeric	sentiment_vader	21	0.9979002	NA	NA	NA	NA	NA	NA	NA	NA	0.6600209	0.4565488	-0.982	0.601	0.859	0.944	1	▁▁▁▁▇

#histogram of sentiment
hist(df_vader$sentiment_vader)

#see what goes into compound sentiment breakdown
df_vader_breakdown <- df_vader %>%
  dplyr::group_by(review_id) %>%
  dplyr::mutate(vader_output = vader_df(text)) %>% #generate vader output
  dplyr::mutate(prop_positive = vader_output$pos, #create columns showing proportion of each review that are positive, negative, and neutral
         prop_neutral = vader_output$neu,
         prop_negative = vader_output$neg) %>%
  select(review_id, prop_positive, prop_neutral, prop_negative) %>%
  dplyr::ungroup() %>%
  pivot_longer(cols = c(prop_positive, prop_neutral, prop_negative), names_to = "sentiment_vader", values_to = "proportion") %>%
  dplyr::mutate(sentiment_vader = factor(sentiment_vader, levels = c("prop_negative","prop_neutral","prop_positive"), labels = c("negative", "neutral", "positive"))) %>% #relabeling negative, neutral, and positive proportions
  dplyr::filter(!(proportion < 0 | proportion > 1)) %>% #filter out miscalculated proportions
    drop_na(proportion)

#plot proportion of each review that is negative, positive, and neutral
df_vader_breakdown %>%
  ggplot(aes(x = as.factor(review_id), y = proportion, fill = as.factor(sentiment_vader))) +
  geom_bar(stat = "identity") +
  scale_fill_manual(values = c("red3", "orange2", "gold")) +
  labs(title = "Proportions of negative, neutral, and positive VADER sentiment across reviews",
       subtitle = "These data underlie the compound VADER sentiment values.",
       x = "Review",
       y = "Proportion",
       fill = "Sentiment (VADER)") +
  ylim(0,1) +
  coord_flip() +
  theme_classic() +
  theme(legend.position = "bottom") +
  theme(axis.text.x=element_blank()) 

Takeaways: Reviews are mostly positive or neutral in their language used according to VADER.

#plot association between vader linguistic sentiment and review score
df_vader %>%
  drop_na(sentiment_vader, score) %>%
  ggplot(aes(y = as.numeric(sentiment_vader), x = as.factor(score), color = as.factor(score), fill = as.factor(score))) +
  geom_jitter(size = 1, color = "gray") +
  geom_violin(alpha = 0.7) +
  scale_fill_manual(values = MetBrew_Egypt) +
      scale_color_manual(values = MetBrew_Egypt) +
  geom_boxplot(width = 0.1, color = "white") +
  labs(x = "Review Score", y = "Review linguistic sentiment (VADER)") +
  theme_minimal() +
  theme(legend.position = "none") 

Takeaways: More positive linguistic sentiment via VADER is directionally associated with much better review scores.

#plot association between vader linguistic sentiment and review length
df_vader %>%
  subset(review_wordcount <= (84.64+(3*82.70))) %>% #subset to remove review length outliers more than 3 SD from average review length
  drop_na(sentiment_vader, review_wordcount) %>%
  ggplot(aes(y = as.numeric(sentiment_vader), x = as.numeric(review_wordcount))) +
  geom_jitter(size = 1, color = "gray") +
  geom_smooth(method = "loess") +
  labs(x = "Review Wordcount", y = "Review linguistic sentiment (VADER)") +
  theme_minimal() +
  theme(legend.position = "none") 

## `geom_smooth()` using formula = 'y ~ x'

Takeaways: Reviews have slightly more positive linguistic sentiment via VADER as they get longer, but are fairly positive in linguistic sentiment to begin with.

Method comparison: Using a sentiment lexicon (AFINN; Nielsen, 2011)

#select sentiment database - afinn
AFINN <- tidytext:: get_sentiments("afinn")

#join sentiment values to review data by word matches
df_AFINN <- df_words %>%
  inner_join(AFINN, by = c("word")) %>%
  rename("sentiment_AFINN" = "value")

#examine sentiment
knitr::kable(skim(df_AFINN))

skim_type	skim_variable	n_missing	complete_rate	character.min	character.max	character.empty	character.n_unique	character.whitespace	factor.ordered	factor.n_unique	factor.top_counts	numeric.mean	numeric.sd	numeric.p0	numeric.p25	numeric.p50	numeric.p75	numeric.p100	numeric.hist
character	score	0	1	3	3	0	5	0	NA	NA	NA	NA	NA	NA	NA	NA	NA	NA	NA
character	highly_reviewed	0	1	15	19	0	2	0	NA	NA	NA	NA	NA	NA	NA	NA	NA	NA	NA
character	word	0	1	2	17	0	779	0	NA	NA	NA	NA	NA	NA	NA	NA	NA	NA	NA
factor	review_id	0	1	NA	NA	NA	NA	NA	FALSE	9777	363: 65, 539: 65, 484: 57, 360: 55	NA	NA	NA	NA	NA	NA	NA	NA
factor	productId	0	1	NA	NA	NA	NA	NA	FALSE	3465	853: 502, 756: 458, 915: 449, 602: 424	NA	NA	NA	NA	NA	NA	NA	NA
factor	userId	0	1	NA	NA	NA	NA	NA	FALSE	9070	590: 215, 412: 130, 135: 120, 527: 84	NA	NA	NA	NA	NA	NA	NA	NA
numeric	n_reviews	0	1	NA	NA	NA	NA	NA	NA	NA	NA	15.666661	19.184105	1	2	6	23	82	▇▂▁▁▁
numeric	review_wordcount	0	1	NA	NA	NA	NA	NA	NA	NA	NA	146.064745	148.515108	10	56	100	181	1318	▇▁▁▁▁
numeric	sentiment_AFINN	0	1	NA	NA	NA	NA	NA	NA	NA	NA	1.481144	1.792796	-4	1	2	3	5	▁▂▂▇▁

#histogram of sentiment
hist(df_AFINN$sentiment_AFINN)

#see distribution of average AFINN sentiment across reviews
df_AFINN %>%
  dplyr::group_by(review_id) %>%
  dplyr::mutate(average_sentiment_AFINN = mean(sentiment_AFINN),
         primary_sent_direction = ifelse(average_sentiment_AFINN > 0, "positive", "negative")) %>% #create columns for average sentiment per review and average direction positive or negative
  dplyr::ungroup() %>%
  select(c(review_id, average_sentiment_AFINN, primary_sent_direction)) %>%
  distinct() %>%
  ggplot(aes(x = average_sentiment_AFINN, y = reorder(as.factor(review_id), average_sentiment_AFINN), color = as.factor(primary_sent_direction))) + #plot average sentiment scores with color by average direction
  geom_col() +
  scale_color_manual(values = c("red3", "green3")) +
  labs(title = "Average AFINN sentiment distribution across reviews",
       y = "Review",
       x = "Average AFINN sentiment",
       fill = "Sentiment (VADER)") +
  theme_classic() +
    theme(axis.text.y=element_blank()) +
  theme(legend.position = "none")

## Ignoring unknown labels:
## • fill : "Sentiment (VADER)"

Takeaways: Reviews are much more positive than negative in their linguistic sentiment according to AFINN.

#plot association between AFINN sentiment and review score
df_AFINN %>%
  dplyr::group_by(review_id) %>%
  dplyr::mutate(average_sentiment_AFINN = mean(sentiment_AFINN)) %>%
  dplyr::ungroup() %>%
  select(c(review_id, average_sentiment_AFINN, score)) %>%
  distinct() %>%
  ggplot(aes(y = as.numeric(average_sentiment_AFINN), x = as.factor(score), color = as.factor(score), fill = as.factor(score))) +
  geom_jitter(size = 1, color = "gray") +
  geom_violin(alpha = 0.7) +
  scale_fill_manual(values = MetBrew_Egypt) +
      scale_color_manual(values = MetBrew_Egypt) +
  geom_boxplot(width = 0.1, color = "white") +
  labs(x = "Review Score", y = "Average linguistic sentiment (AFINN)") +
  theme_minimal() +
  theme(legend.position = "none") 

Takeaways: More positive linguistic sentiment via AFINN is directionally associated with slightly better review scores.

#plot association between AFINN sentiment and review length
df_AFINN %>%
  subset(review_wordcount <= (84.64+(3*82.70))) %>% #subset to remove review length outliers more than 3 SD from average review length
  dplyr::group_by(review_id) %>%
  dplyr::mutate(average_sentiment_AFINN = mean(sentiment_AFINN)) %>%
  dplyr::ungroup() %>%
  drop_na(average_sentiment_AFINN, review_wordcount) %>%
  ggplot(aes(y = as.numeric(average_sentiment_AFINN), x = as.numeric(review_wordcount))) +
  geom_jitter(size = 1, color = "gray") +
  geom_smooth(method = "loess") +
  labs(x = "Review Wordcount", y = "Review linguistic sentiment (AFINN)") +
  theme_minimal() +
  theme(legend.position = "none") 

## `geom_smooth()` using formula = 'y ~ x'

Takeaways: Reviews have similar linguistic sentiment via AFINN regardless of length.

What did we learn?

In this workshop, we:

• Set up a text analysis plan in R

• Learned about how and why to use different text cleaning steps

• Examined common experiences and painpoints discussed in a random subset of Amazon fine foods reviews (McAuley & Leskovec, 2013) using basic text analyses and visualizations, including:

  • Evaluation of most frequent words, bigrams, and associations between words in all reviews and those with high vs. low stars

  • Summarization of themes using topic modeling (LDA)

  • Examination of linguistic sentiment via sentiment analysis

    • Comparison of multiple methods to highlight the importance of selecting use-case-appropriate methods

• Developed meaningful business insights from a subset of unstructured review text data

Remember to carefully evaluate data sources and quality, clean text data in line with best practices, choose analyses that are best suited to your use case, and interpret results in context and keeping in mind that they are often observational.

Happy text analyzing!

References & Additional Resources

Additional resources

• UPenn Library Guide to text analysis
• Text Mining with R book by Julia Silge and David Robinson
• Supervised Machine Learning for Text Analysis in R book by Julia Silge and Emil Hvitfeldt
• An Introduction to Text Processing and Analysis with R vignette by Michael Clark
• Ethics Sheet for Automatic Emotion Recognition and Sentiment Analysis (2022) by Saif Mohammad
• stringr package Regular Expressions vignette

References

McAuley, J. & Leskovec, J. (2013). From amateurs to connoisseurs: modeling the evolution of user expertise through online reviews. WWW, 2013.

Porter, M. F. 2001. “Snowball: A Language for Stemming Algorithms.” https://snowballstem.org.

Blei, D. M., Ng, A. Y., & Jordan, M. I. (2003). Latent Dirichlet Allocation. Journal Of Machine Learning Research, 3(4/5), 993-1022. https://www.jmlr.org/papers/volume3/blei03a/blei03a.pdf

Hutto, C., & Gilbert, E. (2014). VADER: A Parsimonious Rule-Based Model for Sentiment Analysis of Social Media Text. Proceedings of the
International AAAI Conference on Web and Social Media, 8(1), 216-225. https://doi.org/10.1609/icwsm.v8i1.14550

Nielsen, F. Å. (2011) A new ANEW: Evaluation of a word list for sentiment analysis in microblogs. Proceedings of the ESWC2011 Workshop on ‘Making Sense of Microposts’: Big things come in small packages 718 in CEUR Workshop Proceedings 93-98. 2011 May. http://arxiv.org/abs/1103.2903

R Core Team (2025). R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria. https://www.R-project.org/.

Wickham H, Averick M, Bryan J, Chang W, McGowan LD, François R, Grolemund G, Hayes A, Henry L, Hester J, Kuhn M, Pedersen TL, Miller E, Bache SM, Müller K, Ooms J, Robinson D, Seidel DP, Spinu V, Takahashi K, Vaughan D, Wilke C, Woo K, Yutani H (2019). “Welcome to the tidyverse.” Journal of Open Source Software, 4(43), 1686. doi:10.21105/joss.01686 https://doi.org/10.21105/joss.01686.

Wickham H (2023). stringr: Simple, Consistent Wrappers for Common String Operations. doi:10.32614/CRAN.package.stringr https://doi.org/10.32614/CRAN.package.stringr, R package version 1.5.1, https://CRAN.R-project.org/package=stringr.

Waring E, Quinn M, McNamara A, Arino de la Rubia E, Zhu H, Ellis S (2022). skimr: Compact and Flexible Summaries of Data. doi:10.32614/CRAN.package.skimr https://doi.org/10.32614/CRAN.package.skimr, R package version 2.1.5, https://CRAN.R-project.org/package=skimr.

Silge J, Robinson D (2016). “tidytext: Text Mining and Analysis Using Tidy Data Principles in R.” JOSS, 1(3). doi:10.21105/joss.00037 https://doi.org/10.21105/joss.00037, http://dx.doi.org/10.21105/joss.00037.

Rinker, T. W. (2018). textclean: Text Cleaning Tools version 0.9.3. Buffalo, New York. https://github.com/trinker/textclean

Fellows I (2018). wordcloud: Word Clouds. doi:10.32614/CRAN.package.wordcloud https://doi.org/10.32614/CRAN.package.wordcloud, R package version 2.6, https://CRAN.R-project.org/package=wordcloud

Feinerer I, Hornik K (2025). tm: Text Mining Package. doi:10.32614/CRAN.package.tm https://doi.org/10.32614/CRAN.package.tm, R package version 0.7-16, https://CRAN.R-project.org/package=tm.

Roehrick K (2020). vader: Valence Aware Dictionary and sEntiment Reasoner (VADER). doi:10.32614/CRAN.package.vader https://doi.org/10.32614/CRAN.package.vader, R package version 0.2.1, https://CRAN.R-project.org/package=vader.

Grün B, Hornik K (2011). “topicmodels: An R Package for Fitting Topic Models.” Journal of Statistical Software, 40(13), 1–30. doi:10.18637/jss.v040.i13.

Mills BR (2022). MetBrewer: Color Palettes Inspired by Works at the Metropolitan Museum of Art. doi:10.32614/CRAN.package.MetBrewer https://doi.org/10.32614/CRAN.package.MetBrewer, R package version 0.2.0, https://CRAN.R-project.org/package=MetBrewer.

Posit team (2025). RStudio: Integrated Development Environment for R. Posit Software, PBC, Boston, MA. URL http://www.posit.co/.

## ─ Session info ───────────────────────────────────────────────────────────────
##  setting  value
##  version  R version 4.5.0 (2025-04-11)
##  os       macOS Sequoia 15.6
##  system   aarch64, darwin20
##  ui       X11
##  language (EN)
##  collate  en_US.UTF-8
##  ctype    en_US.UTF-8
##  tz       America/New_York
##  date     2025-12-12
##  pandoc   3.4 @ /Applications/RStudio.app/Contents/Resources/app/quarto/bin/tools/aarch64/ (via rmarkdown)
##  quarto   1.6.42 @ /Applications/RStudio.app/Contents/Resources/app/quarto/bin/quarto
## 
## ─ Packages ───────────────────────────────────────────────────────────────────
##  package        * version date (UTC) lib source
##  base64enc        0.1-3   2015-07-28 [1] CRAN (R 4.5.0)
##  bslib            0.9.0   2025-01-30 [1] CRAN (R 4.5.0)
##  cachem           1.1.0   2024-05-16 [1] CRAN (R 4.5.0)
##  cli              3.6.5   2025-04-23 [1] CRAN (R 4.5.0)
##  data.table       1.17.2  2025-05-12 [1] CRAN (R 4.5.0)
##  dichromat        2.0-0.1 2022-05-02 [1] CRAN (R 4.5.0)
##  digest           0.6.37  2024-08-19 [1] CRAN (R 4.5.0)
##  dplyr          * 1.1.4   2023-11-17 [1] CRAN (R 4.5.0)
##  evaluate         1.0.3   2025-01-10 [1] CRAN (R 4.5.0)
##  farver           2.1.2   2024-05-13 [1] CRAN (R 4.5.0)
##  fastmap          1.2.0   2024-05-15 [1] CRAN (R 4.5.0)
##  forcats        * 1.0.0   2023-01-29 [1] CRAN (R 4.5.0)
##  fs               1.6.6   2025-04-12 [1] CRAN (R 4.5.0)
##  generics         0.1.4   2025-05-09 [1] CRAN (R 4.5.0)
##  ggplot2        * 4.0.0   2025-09-11 [1] CRAN (R 4.5.0)
##  glue             1.8.0   2024-09-30 [1] CRAN (R 4.5.0)
##  gtable           0.3.6   2024-10-25 [1] CRAN (R 4.5.0)
##  hms              1.1.3   2023-03-21 [1] CRAN (R 4.5.0)
##  htmltools        0.5.8.1 2024-04-04 [1] CRAN (R 4.5.0)
##  janeaustenr      1.0.0   2022-08-26 [1] CRAN (R 4.5.0)
##  jquerylib        0.1.4   2021-04-26 [1] CRAN (R 4.5.0)
##  jsonlite         2.0.0   2025-03-27 [1] CRAN (R 4.5.0)
##  knitr            1.50    2025-03-16 [1] CRAN (R 4.5.0)
##  koRpus           0.13-8  2021-05-17 [1] CRAN (R 4.5.0)
##  koRpus.lang.en   0.1-4   2020-10-24 [1] CRAN (R 4.5.0)
##  labeling         0.4.3   2023-08-29 [1] CRAN (R 4.5.0)
##  lattice          0.22-6  2024-03-20 [1] CRAN (R 4.5.0)
##  lexicon          1.2.1   2019-03-21 [1] CRAN (R 4.5.0)
##  lifecycle        1.0.4   2023-11-07 [1] CRAN (R 4.5.0)
##  lubridate      * 1.9.4   2024-12-08 [1] CRAN (R 4.5.0)
##  magrittr         2.0.3   2022-03-30 [1] CRAN (R 4.5.0)
##  Matrix           1.7-3   2025-03-11 [1] CRAN (R 4.5.0)
##  MetBrewer      * 0.2.0   2022-03-21 [1] CRAN (R 4.5.0)
##  mgcv             1.9-1   2023-12-21 [1] CRAN (R 4.5.0)
##  mnormt           2.1.1   2022-09-26 [1] CRAN (R 4.5.0)
##  modeltools       0.2-24  2025-05-02 [1] CRAN (R 4.5.0)
##  nlme             3.1-168 2025-03-31 [1] CRAN (R 4.5.0)
##  NLP            * 0.3-2   2024-11-20 [1] CRAN (R 4.5.0)
##  pillar           1.11.0  2025-07-04 [1] CRAN (R 4.5.0)
##  pkgconfig        2.0.3   2019-09-22 [1] CRAN (R 4.5.0)
##  plyr             1.8.9   2023-10-02 [1] CRAN (R 4.5.0)
##  psych          * 2.5.3   2025-03-21 [1] CRAN (R 4.5.0)
##  purrr          * 1.0.4   2025-02-05 [1] CRAN (R 4.5.0)
##  qdapRegex        0.7.10  2025-03-24 [1] CRAN (R 4.5.0)
##  R6               2.6.1   2025-02-15 [1] CRAN (R 4.5.0)
##  rappdirs         0.3.3   2021-01-31 [1] CRAN (R 4.5.0)
##  RColorBrewer   * 1.1-3   2022-04-03 [1] CRAN (R 4.5.0)
##  Rcpp             1.0.14  2025-01-12 [1] CRAN (R 4.5.0)
##  readr          * 2.1.5   2024-01-10 [1] CRAN (R 4.5.0)
##  repr             1.1.7   2024-03-22 [1] CRAN (R 4.5.0)
##  reshape2         1.4.4   2020-04-09 [1] CRAN (R 4.5.0)
##  rlang            1.1.6   2025-04-11 [1] CRAN (R 4.5.0)
##  rmarkdown        2.29    2024-11-04 [1] CRAN (R 4.5.0)
##  rstudioapi       0.17.1  2024-10-22 [1] CRAN (R 4.5.0)
##  S7               0.2.0   2024-11-07 [1] CRAN (R 4.5.0)
##  sass             0.4.10  2025-04-11 [1] CRAN (R 4.5.0)
##  scales           1.4.0   2025-04-24 [1] CRAN (R 4.5.0)
##  sessioninfo      1.2.3   2025-02-05 [1] CRAN (R 4.5.0)
##  skimr          * 2.1.5   2022-12-23 [1] CRAN (R 4.5.0)
##  slam             0.1-55  2024-11-13 [1] CRAN (R 4.5.0)
##  SnowballC        0.7.1   2023-04-25 [1] CRAN (R 4.5.0)
##  stringi          1.8.7   2025-03-27 [1] CRAN (R 4.5.0)
##  stringr        * 1.5.1   2023-11-14 [1] CRAN (R 4.5.0)
##  sylly            0.1-6   2020-09-20 [1] CRAN (R 4.5.0)
##  sylly.en         0.1-3   2018-03-19 [1] CRAN (R 4.5.0)
##  syuzhet          1.0.7   2023-08-11 [1] CRAN (R 4.5.0)
##  textclean      * 0.9.3   2018-07-23 [1] CRAN (R 4.5.0)
##  textdata         0.4.5   2024-05-28 [1] CRAN (R 4.5.0)
##  textshape        1.7.5   2024-04-01 [1] CRAN (R 4.5.0)
##  textstem         0.1.4   2018-04-09 [1] CRAN (R 4.5.0)
##  tibble         * 3.3.0   2025-06-08 [1] CRAN (R 4.5.0)
##  tidyr          * 1.3.1   2024-01-24 [1] CRAN (R 4.5.0)
##  tidyselect       1.2.1   2024-03-11 [1] CRAN (R 4.5.0)
##  tidytext       * 0.4.2   2024-04-10 [1] CRAN (R 4.5.0)
##  tidyverse      * 2.0.0   2023-02-22 [1] CRAN (R 4.5.0)
##  timechange       0.3.0   2024-01-18 [1] CRAN (R 4.5.0)
##  tm             * 0.7-16  2025-02-19 [1] CRAN (R 4.5.0)
##  tokenizers       0.3.0   2022-12-22 [1] CRAN (R 4.5.0)
##  topicmodels    * 0.2-17  2024-08-14 [1] CRAN (R 4.5.0)
##  tzdb             0.5.0   2025-03-15 [1] CRAN (R 4.5.0)
##  vader          * 0.2.1   2020-09-07 [1] CRAN (R 4.5.0)
##  vctrs            0.6.5   2023-12-01 [1] CRAN (R 4.5.0)
##  withr            3.0.2   2024-10-28 [1] CRAN (R 4.5.0)
##  wordcloud      * 2.6     2018-08-24 [1] CRAN (R 4.5.0)
##  xfun             0.52    2025-04-02 [1] CRAN (R 4.5.0)
##  xml2             1.3.8   2025-03-14 [1] CRAN (R 4.5.0)
##  yaml             2.3.10  2024-07-26 [1] CRAN (R 4.5.0)
## 
##  [1] /Library/Frameworks/R.framework/Versions/4.5-arm64/Resources/library
##  * ── Packages attached to the search path.
## 
## ──────────────────────────────────────────────────────────────────────────────